Segmentation system and its evaluation for gray scale coin documents

The numismatics community undergoes an important change toward a digital archiving of coin trades. Transactions of historical coins are tracked in many catalogues and the digitization of such content is of great interest for augmenting experts work in estimating a coin value and preventing illegal sales. In order to recognize and index coin pictures in transactions archives, a segmentation step is required. We propose a system which enables the automatic segmentation of coins and labels in digitized catalogues. As opposed to state of the art approaches, our system can extract coin images of any shape and handle multiple coins per page, as well as text labels. We also present in detail the evaluation metric that we used to evaluate the performance of the proposed system. The evaluation of the proposed system on seven books (230 pages) shows promising results: over 99% of coins and 90% of labels are correctly segmented.